Connecting an AI assistant to your support inbox with MCP

AI assistants are becoming part of everyday work. Support leads ask them to summarize the week, engineers ask them to find the conversation where a bug was first reported, and managers ask them which topics are growing. To do any of that, the assistant needs safe access to your support data.

The Model Context Protocol, or MCP, is an open standard for exactly this. This post explains what it is, how it applies to a support inbox and how to set it up safely.

Update, November 2026: the Inboxbee MCP server is available. Open Developers › MCP server in the dashboard to create a token and connect Claude, ChatGPT or any client that supports Streamable HTTP.

What MCP is

MCP is a common language between AI applications and the tools they use. An MCP server exposes a set of tools, such as "search conversations" or "add a note". An MCP client, for example a desktop AI assistant or an internal agent, connects to the server, discovers those tools and calls them when it needs to.

Before MCP, every AI tool needed its own custom integration for every product. With MCP, a product publishes one server and any compatible assistant can use it.

Why it matters for support

A support inbox holds some of the most useful information a business has: what customers ask, what confuses them, what breaks. MCP lets an assistant work with that information directly.

Useful things to ask an assistant connected to your inbox:

• "Summarize this week's conversations about shipping delays."
• "Find conversations where customers mention the new pricing page."
• "Draft a reply to the oldest open conversation in the billing team."
• "Which saved replies were used most this month?"

Permissions come first

Giving an AI assistant access to customer conversations is a serious decision. A good MCP setup answers three questions.

Who is connecting? Each assistant gets its own MCP token, created by a teammate in the Developers area. Replies the assistant sends appear as that teammate, so customers and your team always see a real person behind them. Give each assistant its own named token so you can tell them apart and revoke one without touching the others.

What can it do? Tools are grouped by risk, and the token decides which ones the assistant sees:

• Read only: search and read conversations, contacts and help articles.
• Read and write: also reply to customers, assign or close conversations and add internal notes.

A sensible start is read-only. Test tokens only see your sandbox, so you can experiment without touching real customers.

What happened? Every call is logged under Developers › Logs with the token's name, so you can audit what an assistant did and when.

Setting it up

The setup takes a few minutes:

1. Open Developers in the dashboard and choose MCP server.
2. Click Create token, name it after the assistant (for example "Claude") and choose Read only.
3. Copy the setup for your client. The screen has ready-made snippets for Claude, Claude Desktop, Claude Code, ChatGPT, Cursor and VS Code. For a client that takes a URL and headers, it looks like this:

{
  "mcpServers": {
    "livechat": {
      "url": "https://<your-api-address>/v1/mcp",
      "headers": { "Authorization": "Bearer <your MCP token>" }
    }
  }
}

1. For Claude or ChatGPT connectors, which take only a URL, paste the URL shown in the dashboard that already contains the token. Treat that URL like a password.
2. Try a read-only request first, such as summarizing yesterday's conversations.

Turn on Test mode in Developers to create a test token while you experiment, so nothing touches real customers.

Good habits

• Start read-only. Create a read-and-write token only when you have a clear use for it.
• One token per assistant, so you can revoke one without breaking the others.
• Review the logs weekly for the first month.
• Revoke tokens you no longer use.
• Tell your team which assistants are connected and what they can do.

MCP, the API and webhooks

MCP is not a replacement for the REST API. Use the API and webhooks for fixed, predictable automations, such as syncing contacts to your CRM. Use MCP when a person is asking an assistant open-ended questions or wants help with one-off tasks.

MCP also works the other way round: with AI Actions, your Inboxbee AI agent can call your own systems through their MCP servers or simple HTTP actions, and it asks the customer to confirm before anything changes. See the MCP integration page for the details.